Analysis of object detection accuracy based on the density of 3D point clouds for deep learning-based shipyard datasets

Abstract

3D point clouds are a crucial data format for accurately capturing geometric information in large-scale industrial environments such as shipyards. Deep learning-based object detection technology using 3D point clouds enables automated production management and process optimization. However, the large volume characteristic of 3D point clouds remains a challenge due to the resources and time required for data processing and dataset construction. The large volume of 3D point clouds leads to excessive computational costs, storage demands, and time consumption during dataset construction and training. Therefore, it is necessary to appropriately reduce the dataset size for efficient utilization while ensuring object detection performance. This necessitates a study on dataset downsampling strategies that maintain optimal density and detection accuracy. In this study, an experimental dataset similar to the S3DIS (Stanford Large-Scale 3D Indoor Spaces) dataset was constructed. The density of the 3D point clouds was adjusted in five levels by reducing points per unit area by 20% increments. These datasets were applied to a deep learning architecture to analyze object detection accuracy. Subsequently, the findings were applied to a shipyard dataset to streamline large volume point clouds and evaluate detection performance, thereby assessing their practical applicability. The results demonstrated that reducing the experimental dataset density to approximately 20% still maintained object detection accuracy of around 95% IoU for key objects. This indicates that lightweight datasets can reduce processing resources and costs while preserving detection performance. Additionally, applying the approach to real shipyard datasets revealed that object detection was feasible with reduced data (approximately 4.6% of the raw data). This study provides a practical framework for constructing efficient deep learning models for object detection by downsampling datasets in large-scale industrial environments like shipyards. It is expected to contribute to the establishment of automated data management systems for production management and process efficiency enhancement. Further analysis is required to evaluate performance at extreme low densities (below 20%). Moreover, while this study employed simple downsampling techniques, future work should explore the performance of various downsampling methods to optimize detection accuracy.